Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of...

9
Achiever or Explorer? Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick [email protected] [email protected] Technische Universität Dresden Dresden, Germany Gregor Blichmann Martin Voigt [email protected] [email protected] AI4BD Deutschland GmbH Dresden, Germany Raimund Dachselt [email protected] Centre for Tactile Internet with Human-in-the-Loop (CeTI), Technische Universität Dresden Dresden, Germany Figure 1: Two screenshots of the gamied annotation tool for the creation of training data for machine learning processes: a passed level (left) and an example annotation task for handwriting recognition (right) ABSTRACT The development of articial intelligence, e. g., for Computer Vision, through supervised learning requires the input of large amounts of annotated or labeled data objects as training data. The creation of high-quality training data is usually done manually which can be repetitive and tiring. Gamication, the use of game elements in a non-game context, is one method to make tedious tasks more interesting. This paper proposes a multi-step process for gamifying the manual creation of training data for machine learning purposes. We choose a user-adapted approach based on the results of a preced- ing user study with the target group (employees of an AI software development company) which helped us to identify annotation use cases and the users’ player characteristics. The resulting concept includes levels of increasing diculty, tutorials, progress indicators and a narrative built around a robot character which at the same time is a user assistant. The implemented prototype is an extension of the company’s existing annotation tool and serves as a basis for further observations. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. Request permissions from [email protected]. MuC’20, September 6–9, 2020, Magdeburg, Germany © 2020 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-7540-5/20/09. . . $15.00 https://doi.org/10.1145/3404983.3405519 CCS CONCEPTS Software and its engineering ! Interactive games; Com- puting methodologies ! Object recognition; Information systems ! Enterprise applications. KEYWORDS gamication, object labeling, training data, machine learning ACM Reference Format: Sarah Alaghbari, Annett Mitschick, Gregor Blichmann, Martin Voigt, and Raimund Dachselt. 2020. Achiever or Explorer? Gamifying the Creation Process of Training Data for Machine Learning. In Mensch und Computer 2020 (MuC’20), September 6–9, 2020, Magdeburg, Germany. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3404983.3405519 1 INTRODUCTION Articial intelligence (AI) is becoming increasingly important. For the development of articial intelligence, human intelligence is still necessary, especially regarding supervised learning which entails that a machine is trained with labeled data. The training process mimics a human learning process, deriving patterns and creating a model. The creation of necessary labels is usually performed with the aid of humans. Due to the necessary amount of training data the creation process is typically highly repetitive and quickly turns into a rather unexciting, demotivating task for the annotator. A task that is repetitive and tedious turns out to be the ideal use case for applying gamication [19]. Gamication itself is dened as the use of game elements in a non-game context [8], aiming Session 3: VR + Games and Learning 173

Transcript of Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of...

Page 1: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

Achiever or Explorer? Gamifying the Creation Process ofTraining Data for Machine Learning

Sarah AlaghbariAnnett Mitschick

[email protected]@tu-dresden.deTechnische Universität Dresden

Dresden, Germany

Gregor BlichmannMartin Voigt

[email protected]@ai4bd.com

AI4BD Deutschland GmbHDresden, Germany

Raimund [email protected] for Tactile Internet withHuman-in-the-Loop (CeTI),

Technische Universität DresdenDresden, Germany

Figure 1: Two screenshots of the gami�ed annotation tool for the creation of training data for machine learning processes: apassed level (left) and an example annotation task for handwriting recognition (right)

ABSTRACTThe development of arti�cial intelligence, e. g., for Computer Vision,through supervised learning requires the input of large amountsof annotated or labeled data objects as training data. The creationof high-quality training data is usually done manually which canbe repetitive and tiring. Gami�cation, the use of game elementsin a non-game context, is one method to make tedious tasks moreinteresting. This paper proposes a multi-step process for gamifyingthe manual creation of training data for machine learning purposes.We choose a user-adapted approach based on the results of a preced-ing user study with the target group (employees of an AI softwaredevelopment company) which helped us to identify annotation usecases and the users’ player characteristics. The resulting conceptincludes levels of increasing di�culty, tutorials, progress indicatorsand a narrative built around a robot character which at the sametime is a user assistant. The implemented prototype is an extensionof the company’s existing annotation tool and serves as a basis forfurther observations.

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor pro�t or commercial advantage and that copies bear this notice and the full citationon the �rst page. Copyrights for components of this work owned by others than theauthor(s) must be honored. Abstracting with credit is permitted. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior speci�c permissionand/or a fee. Request permissions from [email protected]’20, September 6–9, 2020, Magdeburg, Germany© 2020 Copyright held by the owner/author(s). Publication rights licensed to ACM.ACM ISBN 978-1-4503-7540-5/20/09. . . $15.00https://doi.org/10.1145/3404983.3405519

CCS CONCEPTS• Software and its engineering ! Interactive games; • Com-puting methodologies ! Object recognition; • Informationsystems ! Enterprise applications.

KEYWORDSgami�cation, object labeling, training data, machine learning

ACM Reference Format:Sarah Alaghbari, Annett Mitschick, Gregor Blichmann, Martin Voigt,and Raimund Dachselt. 2020. Achiever or Explorer? Gamifying the CreationProcess of Training Data for Machine Learning. In Mensch und Computer2020 (MuC’20), September 6–9, 2020, Magdeburg, Germany. ACM, New York,NY, USA, 9 pages. https://doi.org/10.1145/3404983.3405519

1 INTRODUCTIONArti�cial intelligence (AI) is becoming increasingly important. Forthe development of arti�cial intelligence, human intelligence is stillnecessary, especially regarding supervised learning which entailsthat a machine is trained with labeled data. The training processmimics a human learning process, deriving patterns and creating amodel. The creation of necessary labels is usually performed withthe aid of humans. Due to the necessary amount of training datathe creation process is typically highly repetitive and quickly turnsinto a rather unexciting, demotivating task for the annotator.

A task that is repetitive and tedious turns out to be the ideal usecase for applying gami�cation [19]. Gami�cation itself is de�nedas the use of game elements in a non-game context [8], aiming

Session 3: VR + Games and Learning

173

Page 2: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

MuC’20, September 6–9, 2020, Magdeburg, Germany Alaghbari, et al.

for certain psychological outcomes such as motivation, enjoyment,and �ow. Previous research shows that a gami�ed environment fordata annotation has the potential to increase user engagement andgrati�cation [12]. Improved user experience is a goal of gami�cation,as are increased participation, the attraction of a younger audience,optimization of work�ows and increased engagement of users, aswell as immediate feedback for the users on their performance [23].

Gami�cation of company workplaces has just recently gainedin importance – not only for the training but also to encourageemployees in their daily work routine. A tool with well-designedgame elements at the workplace can keep employees motivated toperform their tasks [16]. This paper presents the results of our workaiming at integrating game elements into an existing annotationtool for the creation of training data at the AI product companyAI4BD 1. We describe our multi-step development process, therebylaying the foundation for future user studies to investigate the e�ectof the implemented game elements.

2 RELATEDWORKOver the years a lot of di�erent approaches to de�ning and classify-ing gami�cation have been established. During the research, severalterms came up which are linked or subordinated to gami�cation,such as Serious Games, Games with a purpose, playfulness, gameful-ness and many more. However, the general purpose of gami�cationis to motivate the target group [8], or more precisely "using game-based mechanics, aesthetics and game thinking to engage people,motivate action, promote learning, and solve problems" [13]. Thus,gami�cation does not mean that a stand-alone game is to be added,hoping for an improvement in employee engagement, but insteadto analyze game mechanics and visuals, and select game parts whichmatch the use case.

2.1 Game Mechanics and Player Types2.1.1 Game Mechanics. We use Game Mechanics as the hypernymof Game Dynamics and Game Elements. By Game Dynamics, wedenote the strategies and characteristics of games, but also theneeds, a player wants to have ful�lled. These needs are, for example,the strive for competition, exploration or social interaction. We willregard Game Elements as the actual components found in a game,such as points, leaderboards or avatars. Related literature describesgame elements as "the building blocks that can be applied andcombined to gamify any non-game context" [15]. This distinctionbeing made, it is still possible to map one to the other. Table 1shows several Game Dynamics as well as Elements that triggerthem respectively. For example, the Game Dynamic progression canbe supported by the Game Element levels or progress bar, whichis intuitively understandable as the feeling of advancing can betriggered with new levels being reached or even unlocked, as wellas with a progress bar which is �lling up increasingly. The GameElement points can be regarded as a progression trigger, under theassumption that the number of points is an indicator for the player’splaying skills which means that an increase of points correlateswith improved skill. On the other hand, points can also be usedto satisfy the need for competition. The dominating motivator inthese cases is competence (alternatively called "mastery") [17]. Now,1AI4BD Deutschland GmbH, https://ai4bd.com/

Table 1: Game Dynamics and suitable Game Elements

Game Dynamic Game ElementExploration, Surprise [5] Unlockable [5]Story/Narrative [10] Story, Badge,

Achievement [14]Boundaries [10] Limited Resources [8]Competition [14] Leaderboard, Points [14]Resource Acquisition [14] Achievement [14],

Unlockable [5]Status [14] Leaderboard, Levels [14]Cooperation [14] Teams [14]Transaction [14] Gifting [14]Reward [14] Badge, Achievement [14]Progression [14] Levels, Progress Bar,

Points [14]

Source: The references in the table denote the source used for the mapping betweenGame Dynamic and Game Element.

knowing these elements, one might be tempted to simply pick theones with the greatest appeal and surprise the employees witha generic game layer featuring a leaderboard and random scores.However, this method has its drawbacks and is criticized by [4] whocall it the "one size �ts all" approach. They suggest a focus on thecontext which is aimed to be gami�ed and to consider "speci�c userneeds, goals and values". Therefore, we will follow a user-centereddesign.

2.1.2 Player Types. Evidently, the essence of user-centered designlies within the users which is why it is necessary to get familiar withthe players. Some authors even recommend a thorough personalityanalysis of the users, with aid of personality type models such as theBig Five or The Myers-Briggs type indicator [2, 10]. It is assumedthat knowing a player’s personality traits, gami�cation can be builtaccording to their personal needs and thus make it easier to triggertheir intrinsic motivation. However, the personality type can alsoprovide insight into how prone to certain dangers of gami�cationa user might be. Several ways of classifying players have beenestablished so far. Notably, many of them are based on Bartle’s 1996theory of four main player types [3]. Bartle’s theory was developedbased on the question "What do people want out of a MUD (Multi-User Dungeon)?". The author collected players’ answers and thencategorized them into four main motivations which he turned intoplayer types, meaning classes of users participating in a MUD whoshare common primary goals. First, there are Socializers who aimfor inter-player relationships, empathize with people and enjoyobserving them. The Killers are likewise focused on other playersbut aim for imposing themselves on others by attacking them andwant to win at any cost. The Achiever type is also interested inwinning but less to defeat others rather than for the sake of points-gathering and rising in levels. Lastly, Bartle de�nes the group ofExplorers who enjoy progressive actions, �guring out how thingswork and the discovery of interesting features. As these playermotivations are not mutually exclusive, a real-life player is regardedas a combination of all of these types at di�erent rates, of whichsome are more and others are less dominant.

Session 3: VR + Games and Learning

174

Page 3: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

Achiever or Explorer? Gamifying the Creation of Training Data MuC’20, September 6–9, 2020, Magdeburg, Germany

2.1.3 Conclusion. Despite being initially derived from a multi-player game context, Bartle’s theory is still highly present in today’splayer classi�cations. The names used for the player types can varygreatly. The Explorer type, for example, is also referred to as FreeSpirit or Creator [18], Detective or Navigator [9], depending on theparticular focus. What appeals to these explorative players, is agame that is highly adaptable and satis�es their need for autonomywith elements such as custom avatars and many unlockable items.Nonetheless, a game with such elements can still attract users of theAchiever type who may not willing to spend 30 minutes on choos-ing an out�t. The possibility to skip such decorative steps shouldbe given in their behalf, as well as additional elements that feed theAchievers’ competitive needs, such as a leaderboard. A leaderboardmight, however, demotivate less competitive users. Therefore, gameelements should be selected deliberately and with a lot of attentionto the users to prevent unpredicted and undesired behavior.

2.2 Gami�cation for AnnotationHaving observed gami�cation in a general way, we analyzed exist-ing approaches that use game elements in the context of annotation.We present three examples, their setup, features and how they relateto our use case.

2.2.1 Gamification in video labeling. A game for video annotationwas designed in [21]. They thought out three di�erent game ap-proaches: a label vote game, an entity annotation where users wereasked to assign a certain category to a video segment, a click game,where users had to locate a certain object inside the video and clickon it, and a bounding box game, which asked users to draw a boxaround a speci�c object. The last one was implemented and eval-uated with the aid of 20 persons who had not been in touch withthe data or the use case ever before. A questionnaire was answeredas well, showing that the users liked the game but also agreed thatit got more repetitive and boring with time. Used game elementswere a progress bar, levels, an optional leaderboard and statisticsover experience. The author also mentions the struggle of creatinga level system with increasing di�culty for an annotation use casewhile maintaining the accuracy of the results. In general, the qualityof the labels was not satisfying as the resulting bounding boxeswere inaccurate. Users also stated they were not willing to spendmore time on the tool. Still, the author concludes that a gami�edapproach could be of advantage concerning annotation cost, giventhat a very e�cient and well-thought-out game is developed. Weassume that this is an example of the "one size �ts all" gami�cationapproach as apparently, the gami�cation concept did not regardany adaptive measures towards the user needs. However, this mighthave been caused by time limits. It has to be mentioned as well,that in this case a full game was developed from scratch, instead ofincluding game elements into an existing tool. Also, unlike our usecase, participants were not regularly confronted with an annotationuse case and therefore they did not have any experience with thistask. We do not see the goal of gami�cation in convincing everypossible user, but in the adaptation and improvement of a tool fora certain group of users.

2.2.2 Tags You Don’t Forget: Gamified Tagging of Personal Images.Another approach was created by [20] whose scope was the cre-ation of a game, used to annotate personal photos. Two mobileapplications were developed (one single, one multiplayer) and eval-uated as well as compared to a simple tagging app without anygami�cation. Concerning game elements, the authors mention thatsimple playful elements, for example, acoustic feedback for inter-action, can already be su�cient in order to motivate a user. Thesingle-player app was a simple tagging app, while the multiplayerapp was developed as a Tagger-Guesser-Game. Here, Player B wasshown a photo and had to choose between several tags to guessthe one that Player A had selected. This approach is similar tothe ESP game [22], which uses human aid for image recognition.Assigned labels were rewarded with a point. Correct answers inthe multiplayer app were likewise rewarded with one point, whileone point was lost for a wrong answer. The labels were evaluatedby an expert as being of "good quality". Besides, a questionnairewas answered by the participants to analyze their impressions ofthe game. They stated that the multiplayer app was much moreentertaining, whereas the single-player app helped them memorizethe labels better. The insight we take from this example is thatless is more when it comes to the selection and amount of gameelements.

2.2.3 Crowdsourcing. Lastly, we analyzed crowdsourcing tools,which often include game elements to engage users. Google Crowd-source [11] is a desktop platform as well as a mobile app, whichmakes use of humans to improve Google tools such as Google Pho-tos or Google Translate and can be used by anyone who has aGoogle account. There are di�erent kinds of tasks that can be per-formed, e. g., image labeling, approval of image labels, handwritingrecognition, and translation, which is close to our use case. In con-trast to the two above-mentioned examples, Google Crowdsourceis an established user contribution platform which is not a gamein itself, but includes game elements, like levels, points, badges,and a leaderboard. It simply works by triggering the basic humanneeds [17]: relatedness, since everything is open and visible, auton-omy since there is no pressure and users are free to decide whenor whether they participate, and above all purpose since users getthe feeling of being an active part in the improvement of popularsoftware tools.

Another crowdsourcing-based approach incorporating game el-ements was proposed by Altmeyer et al. [1]. Their goal was to en-courage people to keep track of their expenses using OCR (opticalcharacter recognition) to analyze grocery receipts. The recognitionis trained by crowd input (classifying a given extract of a receiptor categorizing an item). The implemented smartphone applica-tion features achievements, points, and a leaderboard to motivateusers and to increase the amount of user contribution. However,a dilemma of such crowdsourcing-based approaches is the lack ofknowledge about the characteristics of the user group, making itdi�cult to design for speci�c user needs and player types.

2.3 Dangers of Gami�cationAs gami�cation makes use of game elements, it is necessary tokeep in mind that with these elements some of their risks mightalso be adopted. One way to approach this topic has been executed

Session 3: VR + Games and Learning

175

Page 4: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

MuC’20, September 6–9, 2020, Magdeburg, Germany Alaghbari, et al.

by Callan et al. [6], where ten �ctive scenarios of gami�cationare presented which have been wrongly established in businesses.Recurring problems were a lack of goal-orientation, unsuitablegame elements and rewarding, and the danger of revealing toomuch information to the employees which they might attempt touse for their bene�t.

Furthermore, the term addiction is mentioned in this context[2]. Here, however, it is regarded much more as a dependencywhich users might develop if they get used to the presence of gameelements in connectionwith the task to be performed and hence losetheir motivation to perform said task without gami�cation. As notall possible risks and dangers can be foreseen, another importantmeasure is the constant monitoring of user activities, the detectionof abnormalities and suspicious behavior, and a respective adaptionof the system [2]. After all, no ideal gami�cation will be createdfrom the very beginning. Also, no team of employees will stay thesame over a longer period, and with people, preferences and needswill change. A promising long-term solution is the creation of anintelligent, adaptive gami�cation application [2, 4].

3 REQUIREMENT ANALYSISDeterding et al. [7] describe the procedure of developing a success-fully gami�ed tool as a "full circle" process: "from formative userresearch through synthesis, ideation, prototyping, design and usabil-ity testing". Regarding potential risks of gami�cation (e.g. wronglyguided motivation, o�-task behavior, unwanted competition, ad-diction and dependency) that might be adopted into a system, it isnecessary to de�ne a clear goal that is to be achieved to have a focuswhile conceptualizing the approach and productive game elements.Concerning the annotation task, we regard three central metricsthat can be improved: quantity (howmany annotations are created),quality (how good / correct are the annotations), enjoyment (howmuch fun is the annotation task).

In the following, we �rst describe the current annotation processwith the existing tool support, present our �ndings from a surveyamong the employees, and �nally sketch two possible gami�cationconcepts for this use case.

3.1 Current Annotation ProcessThe company’s existing annotation tool is a multi-user web ap-plication prototype which o�ers registered users a sophisticatedannotation environment for collections of images (typically scanneddocuments). The annotation tasks are of four di�erent types:

• handwriting annotation, where annotators are given an im-age of a handwritten sequence of letters and numbers whichthey have to type,

• document classi�cation, where annotators need to classifyparts of a document, e. g., to mark tables inside a form usingsemantic bounding boxes,

• classi�cation, where annotators are asked to identify a givenobject, e. g., if an image contains a number,

• natural language processing (NLP), where annotators areasked to assign semantic meaning to words, for example, tomark all persons in a given text.

Users can see all the collections which are assigned to them,including their annotation state, i. e., how many of the items within

the collection were approved, annotated and refused (i. e., rejectedbecause it was too ambiguous). Selecting a collection, users cansee a grid view of the contained resources, color-coded dependingon their state (grey: open, blue: approved, red: refused). Addition-ally, users can also see the number of annotations that have beencreated for each resource. Selecting a document, users enter theannotation view itself, where they can create an annotation in casethe document’s state is open, or see its state and annotations thatwere created for this resource. In case users are insecure and wishto access annotation guidelines, they either need to navigate to anexternal annotation tutorial or ask their coworkers.

In order to ensure quality, an annotation is being reviewed afterthe creation. Selected users who have the role "reviewer" assigned tothem can access additional features in the annotation view allowingthem to approve or refuse an annotation. The review of handwritingannotations is currently semi-automated by automatically markingan annotation as "approved" if at least two distinct annotators createan annotation with the same value.

3.2 User SurveyWe conducted a user survey among company employees workingas annotators, to get an idea of their characteristics, whether a gam-i�ed approach would appeal to them at all, which game elementswould suit them most, and which should be avoided regarding theaforementioned potential risks.

We adapted the "student model" from the work of Andrade et al.[2], which de�nes �ve attributes of the player: Knowledge, Psychol-ogy, General Behavior, Gamer Pro�le, and Interaction. As GeneralBehavior is focused on personal habits unrelated to the domain, wedecided to omit this due to privacy issues. The Interaction attributeaddresses information about the user activities which is better ob-tained via monitoring and logging (e. g., number of logins, successrate). We also decided to leave this out as it was not our goal toassess individual user activity.

Consequently, we created a questionnaire covering the threeaspects Knowledge (labeling experience), Psychology (personal opin-ions) and Gamer Pro�le (game experience). Twenty company em-ployees participated in the survey (11 of them aged between 24 and30, two younger than 24, four aged between 31 an 40, one olderthan 40, two preferred not to tell their age). The only mandatoryquestion was if they had ever performed an annotation task. Ifthey had, they could answer more follow-up questions referring toannotations. All other questions were voluntary.

3.2.1 Knowledge. When asked about their experience, 18 out ofthe 20 participants stated that they have already performed an-notation tasks for the company, half of them indicated that theyhave been labeling data for more than three months. In a multiple-choice question, we asked the 18 participants who had experiencewith annotation which kinds of labeling tasks they had already per-formed. Document placement (15) and handwriting recognition (14)were the ones that had been performed by most of the annotators,followed by NLP tasks (8) and classi�cation (6).

3.2.2 Psychology. Concerning the psychological aspect, we askedthe annotators to take a position on six moderately provocativestatements, choosing from a Likert scale of �ve di�erent options

Session 3: VR + Games and Learning

176

Page 5: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

Achiever or Explorer? Gamifying the Creation of Training Data MuC’20, September 6–9, 2020, Magdeburg, Germany

of agreement (I agree... not at all (-2) / not quite (-1) / neutral (0) /a bit (1) / a lot (2)). From the number of positive answers (Likertscale values 1 and 2), we derived a percentage for the agreementper statement.

• "I �nd labeling tasks tiresome" (65% agreed, M=0.7, SD=1.117)• "I would like to be able to see how well I am doing in labeling,compared to my coworkers" (55% agreed, M=0.2, SD=1.348)

• "If labeling included game elements, the label results would bebetter" (50% agreed, M=0.4, SD=0.993)

• "If labeling included game elements it would be much morefun" (65% agreed, M=0.9, SD=0.999)

• "I would not like it if others were able to seemy labeling progresson a leaderboard" (45% agreed, M=0.35, SD=1.27)

• "Using game elements at work makes a company seem lessserious" (30% agreed, 55% disagreed, M=-0.65,SD=1.306)

From these results we derived the following conclusions: Whilelabeling is generally considered rather tiring, the score for thestatements encouraging game elements were overall positive andgame elements are agreed upon as a promising tool for makinglabeling more fun without harming the image of the company.However, most of the annotators dislike the idea of being able tosee their labeling progress on a leaderboard, but do not generallyobject to compare themselves with colleagues.

3.2.3 Gamer Profile. As for the gaming habits, we posed questionsregarding the time spent on games, which kinds of games werepreferred as well as which game elements were the most motivatingones. The majority likes digital games, with 60% of them playingthem at least every week, whereas 20% played them at least once amonth and 10% only rarely or not at all, respectively. We added thequestion on real-life games, in case the participants were not keenon digital games, but still liked playing physically. In our group,however, digital games were more popular.

To �gure out which of Bartle’s four main player types [3] wasmost present in the study group, we asked the participants to indi-cate how much they enjoy distinctive game types (like simulationgames, action games, puzzle-based games, etc.), using a �ve-levelLikert scale (not at all(-2) ... a lot (2)). Furthermore, they were alsoasked to rate speci�c game elements (like leaderboards, playingagainst others, rewards, team play, etc.) on a �ve-level Likert scale(very demotivating (-2) ... very motivating (2)). Finally, we also askedabout the dominating motivation to play games at all. We used thecorrelation between preferred game types, preferred game elementsand gaming motivations to create a score for each set of player typecharacteristics per player. The resulting scores are shown in Figure 2.We identi�ed nine participants with predominant characteristics ofan Achiever (points-gathering, rising in levels), six more inclinedto be an Explorer (progressive actions, �nd interesting features),and two showing equal characteristics of both (P9, P15). Thus, agroup tendency towards Achiever and Explorer characteristics wasnotable.

3.2.4 Further Feedback. We also asked the annotators what theydisliked about the current tool and how they would like it to beimproved. From this information, we hoped to be getting someimpressions of possible stimuli for a gami�cation concept. Except

Figure 2: Player type characteristics of each participant. Thescore is based on the answers regarding preferred gametypes, preferred game elements and gaming motivation.Note: P13 and P17 are excluded here as both considered allelements to be "demotivating" (negative score).

statements like "Labeling is boring.", the feedback we got for thisquestion was concerning more technical issues, like the repeateddemand for a more �uid user interaction inside the annotationtool by supporting key shortcuts to reduce the need for mouseinteraction.

3.2.5 Lessons learned. From our survey, we can conclude that themajority of annotators are playing games and are open to the useof game elements inside work tasks. We learned that a way of com-parison of performance is desired, but should provide anonymity,which is also a precaution in terms of the danger Unwanted Compe-tition. A complex narrative, levels with increasing di�culty, as wellas playing with others in a team, but also playing against others andexploration were voted as the most motivating game elements. Thedominating player type characteristics in the group of annotatorshence turned out to be the ones of the Achiever and the Explorertype. Based on these �ndings, we created a gami�ed concept for anannotation tool, described in the following.

3.3 Basic Gami�cation ConceptsWe combined the results of the survey into two di�erent concepts,each containing a selection of the preferred game elements:

(1) I can make a change: Story / narrative, levels, progress bar,badge / reward (Explorer type)

(2) TeamChallenge - Us versus them: Competition between teams,leaderboard with team names, points and achievements(Achiever type)

We presented both basic concepts to the annotators themselves,giving them a chance to give feedback and express their opinionsconcerning the idea of having said game elements inside theirtools. While the competitive approach seemed appealing, the �rstone of building a story around the annotation tasks was preferred.Furthermore, the idea of incorporating a narrative led to muchvaluable creative input on the part of the annotators. Some ideaswhich were named were to divide the big narrative into variouschapters and steps (hence levels) that need to be passed, but alsothat the story - if it is told correctly - could help the annotatorsunderstand what their work was used for and why they were doingit. So subconsciously, they expressed the desire for Purpose (people�nd a task much more intriguing when there is a reason to do it ora greater meaning behind it, which can also be linked to altruismaccording to [17]) .

Session 3: VR + Games and Learning

177

Page 6: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

MuC’20, September 6–9, 2020, Magdeburg, Germany Alaghbari, et al.

Figure 3: The central level overview, with a header showingthe episode tabs and a sidebar with themenu points "my sta-tistics" and "help".

Further, it was mentioned that despite not implementing a teamfeature per se, the tool could still support the group feeling byshowing the general group progress for all annotations. Relatedliterature refers to this aspect as "Perceived visibility", being "relatedto the notion of being noticed by peers and in a position of socialpresence" [2]. On the other hand, concerns were voiced as to howmotivating such a progress bar would be in case there would bea long period without any change or if the model was adapted,causing the progress to decrease. However, the story element leavesa lot of room for adaptations in this case, e. g., negative progresscould be explained using a negative twist inside the narrative itself.This requires a constant monitoring of user performance and athorough player model to detect anomalies and have the systemreact in such a way that user behavior is directed correctly.

4 GAME DESIGN AND IMPLEMENTATIONFrom the discussion with the users, we derived a �nal game conceptthat includes a complex narrative, levels of increasing di�culty,and a progress indicator.

Upon �rst use, the annotators are taken on a tour through thetool where all the central elements are described, and the tasksand the narrative are introduced. Knowing the basic story and themain goal, users get to the level overview (Figure 3) which fromthen on is always going to be the initial screen. Users can switchbetween di�erent episodes that each can be used to tell di�erentstorylines, but also for annotation tasks of di�erent types. Newepisodes can be added by administrators (or authors) at any time,so they do not depend on the user’s progress. Clicking on one level,the users enter the level screen where the di�erent resources whichare assigned to this level are listed (see Figure 1, left). From here,they can choose a resource to annotate (see Figure 1, right). Thefollowing subsections are each dedicated to one game element,describing its design, placement, and function.

4.1 StoryIn order to support intrinsic motivation, a story element for gami�-cation should be linked to the context and the tasks performed bythe company [2]. Hence, we got inspired by Google Crowdsource

[11] where the speech assistant training task is initialized with arobot which introduces the topic to the user. We created the AIuser assistant "Robbie" to ful�ll three jobs: it is a tutor which gives�rst-time users a tour through the tool, it is the "Help" elementof the tool and it is the center of our narrative, telling the storyand giving feedback on the progress. Each episode has one mainstoryline where Robbie faces a struggle that needs to be solvedwith the help of annotations. Examples for these stories are to helpRobbie achieve certain capabilities, such as learning the humanlanguage, which includes learning to "read" (handwriting annota-tion), to "understand" document structures (bounding box tasks)or even linguistic structures (entity annotation). In the followinglevels, these plots can always be reused by asking the user to trainRobbie further in terms of one of the mentioned capabilities. Withthese plots, we aim to support the user need purpose by creatingabstract stories that are related to the real-life use case.

4.2 LevelsInside a level, users can see all of the resources, including the onesannotated by other users. For this reason, there is a �lter bar whichannotators can use to �lter the resources by their state and by "onlymy resources" (switch-toggle button). Additionally, each resourcewhich was annotated or already started by this user has a usericon in the top right corner. In the level head, users can see thelevel’s quest in the center and the number of their current scoreinside this level in the right corner. This score shows the numberof approved or the number of made annotations (depending on thequest) out of the number of annotations needed to pass the level.The quest itself is a narrative element. It asks the user to reach acertain annotation goal which leads to ful�lling a greater purpose(of helping Robbie). The way the quest is phrased and designed isessential for the ful�llment of the gami�cation goal. For now, wedistinguish two basic quest types: "Annotate a certain number ofresources!" (quantity) or "Get a certain number of approvals for yourannotations!" (quality). The main goal of our gami�cation approachis to improve the quality of the annotations. Consequently, forthe most part, our quests will require users to create approvedannotations or combine both quest types ("Annotate x resources andget y approvals!").

4.3 ProgressIn order to keep track of the quest realization, the user’s progressneeds to be visualized. Progress is shown in the level overview,where users can see what percentage of the total resources insidethe level has already been annotated and approved and if theypassed the level or not. Furthermore, the possibility of unlockingnew levels if the user passed a level is another user-speci�c progresselement. Inside each level, there are multiple progress elements: theannotation counter in the level head, the resource colors, the leveltheme image (which changes from greyscale to color upon passinga level), and the level progress bar. In the following sections, weexplain in detail the layout of this progress bar and how we adoptedcolors into the progress concept.

4.3.1 Colors. We use four main colors for encoding progress (Fig-ure 4). According to their state, the background color for a level boxin the level overview and a resource in the level view is determined.

Session 3: VR + Games and Learning

178

Page 7: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

Achiever or Explorer? Gamifying the Creation of Training Data MuC’20, September 6–9, 2020, Magdeburg, Germany

Figure 4: The di�erent colors of resources, depending ontheir states yellow: rework, white: open, green: approved,blue: in review, grey: locked

Figure 5: The progress bar shown inside the level view

Green represents the state passed (level) or approved (resource).White represents the state open and is used for the background ofa level box that contains open resources and of an open resourcewhich still needs to be annotated. Grey denotes the state lockedand is used for levels that are not accessible yet as well as for re-sources that can not be accessed because they are currently beingannotated by another user. A resource has two additional potentialstates: being in review after a user �nished annotating it (whichis shown with blue color), and being in rework if a resource hasbeen reviewed and not approved due to incorrect or insu�cientannotations. The same color palette is used for the progress barinside a level.

4.3.2 Progress Bar. The progress bar contains information on howmany annotations in this level have been approved (green) and howmany annotations are in the review process (blue), in proportionto the whole number of resources. The remaining white spaceof the bar implicitly encodes all open resources inside this level.The green space is additionally divided into several parts, eachone representing a user and their approved annotations. Thesepartitions are sorted from left ("most approved annotations") toright ("least approved annotations") and they are anonymous, so nouser can see which one belongs to whom. They can, however, seewhere their part is located which is highlighted in blue and with auser icon (Figure 5). So, apart from serving as progress information,the bar also serves as an anonymous leaderboard.

4.3.3 Success Notifications. When a level is passed and the nextone gets unlocked, a user will get a success noti�cation. Whereand when exactly it appears, depends on the passing rule. If thequest goal is to create a certain number of annotations only, thesuccess noti�cation will appear inside the annotation tool, showinga happy Robbie celebrating and giving the user the options to goto the next level or proceed to create annotations for the current

Figure 6: A user’s personal statistics page showing achieve-ments, number of annotations created per annotation type,as well as a history chart showing total number of annota-tions over time.

level. If the passing rule demands a certain number of approvals,the moment when a level is passed does not correlate with the an-notation �ow or even with the time when the user is online. In thiscase, the success message will appear in the level overview, with ananimation showing the next level being unlocked. Alternatively, apop-up noti�cation inside the tool can be shown if the user is onlinewhen the required number of annotation approvals is reached.

4.4 StatisticsIn the sidebar, users can access a general help screen by clicking onthe Robbie icon, but also access their personal user statistics. Thisis a feature we proposed due to the fact that users did like the ideaof seeing their progress, also in comparison to others as derivedfrom the results of the user study. In the statistics screen, shown inFigure 6, they can see awards they got, the number of annotationsthey created per annotation type, as well as a chart showing thehistory of their total number of annotations. The ideas for thesecharts are just an initial suggestion and can be adapted to the users’needs.

4.5 TutorialUsers see a tutorial after they click on an open resource if thisrequires annotation of a type which the user is not experiencedin or in case the resource has any kind of exceptional rules whichthe annotator must know. Currently, the company has tutorialsthat are not embedded inside the annotation tool. By providingthis feature, we aim to support the improvement of the annotationquality as it requires users to read the rules and guidelines beforecreating the annotation. The tutorial consists of multiple steps:�rst, the rules are presented, with Robbie as a decorative element,highlighting positive rules (what is recommended, examples for agood dataset) and negative rules (which data should be skipped), asshown in Figure 7. They are followed by a brief training part whereusers get confronted with minimal annotation tasks that test theirunderstanding of the previously presented rules.

Session 3: VR + Games and Learning

179

Page 8: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

MuC’20, September 6–9, 2020, Magdeburg, Germany Alaghbari, et al.

Figure 7: A tutorial page, introducing an annotation taskwith positive and negative rules and recommendations.

4.6 Working PrototypeStarting with low-�delity prototypes, we �nally implemented afully operable Web prototype within the company’s annotationenvironment, based on Angular 9 and NgRx, and using a centralbackend service, which obtains data from a MongoDB database. Inaddition to the former annotation tool, the prototype introducessoftware components to represent episodes, annotation levels andawards as game elements. The prototype is fully operational re-garding real annotation tasks, user authentication and login, aswell as loading and saving annotations per user account. Thus,preexisting and gami�ed annotation tool are ready for use, e. g. tobe compared in long-term A/B tests. Short-term tests have alreadybeen conducted to verify functionality, but these are not yet mean-ingful regarding our four central metrics, quantity, quality andenjoyment. Nevertheless, initial feedback from the company andindividual employees on the results was throughout very positiveand encouraging.

5 CONCLUSION AND FUTUREWORKThis work describes our approach and design process for the gami�-cation of an annotation tool for creating machine learning trainingdata. Unlike a "one size �ts all" approach to gamifying the tool,where game elements are applied without regard to the context ofuse [4], we choose a user-adapted approach which �rst analyzes ex-isting literature for gami�cation and then performs a user researchstudy with twenty employees of the company. The results showthat the employees enjoy gaming in their free time, which supportsthe utilization of a gami�ed tool, and which game preferences theyhave. From the �ndings, we derive an individual gami�cation con-cept with regard to the annotation use case and the employees’player characteristics. Our implemented prototype is a gami�cationapproach for the company’s annotation tools. It serves as a proofof concept that game elements can be easily implemented insidean existing environment.

One aspect we did not cover in this work is how to assess thedi�culty of an annotation task. One approach is to map the com-plexity, hence the e�ort, of the task to the di�culty. On the otherhand, even a less complex task can be of greater e�ort than a com-plex task if the data contains a lot of ambiguousness. Future work

can analyze this problem thoroughly. It might also be helpful tofollow a more thorough approach of user research that considerspsychological aspects, possibly even personality types, and aimsfor a deeper user analysis. Generally, it can be interesting for otherprojects to regard other taxonomies of player types and to performdetailed psychological user research. Besides, we did not evaluateour approach over a long period (over several months), which is es-pecially necessary when a narrative is included which is an elementthat evolves.

During our research, we found a variety of di�erent taxonomiesfor gami�cation. We also noticed that not all terms are used consis-tently by di�erent sources, for example, the terms Game MechanicsandGame Elements. Besides, many approaches to distinguish playertypes exist, which is why we chose to stick with the basic playertype taxonomy by [3]. Researchers with a similar purpose shouldkeep in mind that gami�cation is perceived with skepticism andconcerns by some users as the underlying idea is the manipulationof user behavior. For the prevention of a negative impression, it isrecommended to include the users in the design process by askingfor their opinions and their general game a�nity. We strongly dis-courage any destructive intentions when using gami�cation, forexample, aiming for surveillance of the sta� or a highly competitiveenvironment in the company. Related work also frequently men-tions the importance of transparency and disclosure concerningthe game tool. One possible way to encourage trust is by giving theusers access to information on the reasons for the use of game ele-ments and not leaving them with a wrong feeling of being observedor put under pressure by a gami�ed tool.

ACKNOWLEDGMENTSThis work was partially funded by the German Research Foun-dation (DFG, Deutsche Forschungsgemeinschaft) as part of Ger-many’s Excellence Strategy – EXC 2050/1 – Project ID 390696704– Cluster of Excellence "Centre for Tactile Internet with Human-in-the-Loop" (CeTI) of Technische Universität Dresden, and byDFG grant 389792660 as part of TRR 248 (see https://perspicuous-computing.science). Additionally, the work at AI4BD was fundedby the German Federal Ministry of Education and Research and theresearch project PANDIA (grant number 16SV8396).

REFERENCES[1] Maximilian Altmeyer, Pascal Lessel, and Antonio Krüger. 2016. Expense Control:

A Gami�ed, Semi-Automated, Crowd-Based Approach For Receipt Capturing.In Proceedings of the 21st International Conference on Intelligent User Interfaces(Sonoma, California, USA) (IUI ’16). Association for Computing Machinery, NewYork, NY, USA, 31–42. https://doi.org/10.1145/2856767.2856790

[2] Fernando Andrade, Riichiro Mizoguchi, and Seiji Isotani. 2016. The Bright andDark Sides of Gami�cation. In Intelligent Tutoring Systems, Alessandro Micarelli,John Stamper, and Kitty Panourgia (Eds.). Lecture Notes in Computer Science,Vol. 9684. Springer International Publishing, 1–11. https://doi.org/10.1007/978-3-319-39583-8_17 06.

[3] Richard Bartle. 1996. Hearts, clubs, diamonds, spades: Players who suit MUDs.Journal of MUD research 1, 1 (1996), 19.

[4] Martin Böckle, Isabel Micheel, Markus Bick, and Jasminko Novak. 2018. A DesignFramework for Adaptive Gami�cation Applications. In Proceedings of the 51stHawaii International Conference on System Sciences (Proceedings of the AnnualHawaii International Conference on System Sciences), Tung Bui (Ed.). HawaiiInternational Conference on System Sciences. https://doi.org/10.24251/HICSS.2018.151

[5] Bunchball. 2015. What are Game Mechanics? https://www.bunchball.com/gami�cation/game-mechanics

Session 3: VR + Games and Learning

180

Page 9: Achiever or Explorer? Gamifying the Creation Process of ...Gamifying the Creation Process of Training Data for Machine Learning Sarah Alaghbari Annett Mitschick sarah.alaghbari@tu-dresden.de

Achiever or Explorer? Gamifying the Creation of Training Data MuC’20, September 6–9, 2020, Magdeburg, Germany

[6] Rachel C Callan, Kristina N Bauer, and Richard N Landers. 2015. How to avoidthe dark side of gami�cation: Ten business scenarios and their unintended con-sequences. In Gami�cation in education and business. Springer, 553–568.

[7] Sebastian Deterding, Sta�an Björk, Lennart Nacke, Dan Dixon, and ElizabethLawley. 2013. Designing gami�cation: creating gameful and playful experiences.In Extended Abstracts on Human Factors in Computing Systems. 3263–3266. https://doi.org/10.1145/2468356.2479662 04.

[8] Sebastian Deterding, Dan Dixon, Rilla Khaled, and Lennart Nacke. 2011. FromGame Design Elements to Gamefulness: De�ning "Gami�cation. In Proceedingsof the 15th International Academic MindTrek Conference: Envisioning Future MediaEnvironments (New York, NY, USA) (MindTrek ’11). ACM, 9–15. https://doi.org/10.1145/2181037.2181040

[9] Dominique Mangiatordi. 2018. Gami�cation at work: the 8 PLAYERTYPES. https://www.linkedin.com/pulse/gami�cation-work-8-player-types-dominique-mangiatordi-/

[10] Lauren Ferro, Ste�en Walz, and Stefan Greuter. 2013. Towards personalised,gami�ed systems: An investigation into game design, personality and playertypologies. In Proceedings of The 9th Australasian Conference on InteractiveEntertainment Matters of Life and Death. https://doi.org/10.1145/2513002.251302409.

[11] Google. 2020. Google Crowdsource. https://crowdsource.google.com/[12] Simone Hantke, Tobias Appel, and Björn Schuller. 2018. The Inclusion of Gam-

i�cation Solutions to Enhance User Enjoyment on Crowdsourcing Platforms.In 2018 First Asian Conference on A�ective Computing and Intelligent Interaction(ACII Asia). 1–6. https://doi.org/10.1109/ACIIAsia.2018.8470330 05.

[13] Karl Kapp. 2012. The gami�cation of learning and instruction: Game-based methodsand strategies for training and education. San Francisco, CA: Pfei�er.

[14] Selay Arkün Kocadere and Şeyma Çağlar. 2018. Gami�cation from player typeperspective: A case study. Journal of Educational Technology & Society 21, 3

(2018), 12–22.[15] Janaki Kumar andMario Herger. 2013. Gami�cation atWork: Designing Engaging

Business Software. In Design, User Experience, and Usability. Health, Learning,Playing, Cultural, and Cross-Cultural User Experience (Berlin, Heidelberg), AaronMarcus (Ed.). Springer Berlin Heidelberg, 528–537.

[16] Daria Lopukhina. 2018. How Gami�cation in the Workplace Impacts EmployeeProductivity. https://anadea.info/blog/how-gami�cation-in-the-workplace-impacts-employee-productivity

[17] Andrzej Marczewski. 2013. The Intrinsic Motivation RAMP. https://www.gami�ed.uk/gami�cation-framework/the-intrinsic-motivation-ramp/

[18] Andrzej Marczewski. 2015. Even ninja monkeys like to play: Gami�cation, gamethinking & motivational design. Gami�ed UK. http://gami�ed.uk/user-types/

[19] Jane McGonigal. 2011. Reality is broken. Penguin Press. http://www.loc.gov/catdir/enhancements/fy1107/2010029619-d.html

[20] Nina Runge, Dirk Wenig, Danny Zitzmann, and Rainer Malaka. 2015. Tags YouDon’t Forget: Gami�ed Tagging of Personal Images. In Entertainment Computing- ICEC 2015 (Cham), Konstantinos Chorianopoulos, Monica Divitini, JannickeBaalsrud Hauge, Letizia Jaccheri, and Rainer Malaka (Eds.). Springer InternationalPublishing, 301–314.

[21] Samuel Suikkanen. 2019. Gami�cation in video labeling; Videoidenmerkkaamisen pelillistäminen. http://urn.�/URN:NBN:�:aalto-201906233997

[22] Luis von Ahn and Laura Dabbish. 2004. Labeling images with a computer game.In Proceedings of the 2004 conference on Human factors in computing systems -CHI ’04 (New York, New York, USA), Elizabeth Dykstra-Erickson and ManfredTscheligi (Eds.). ACM Press, 319–326. https://doi.org/10.1145/985692.985733

[23] Lincoln Wood and Torsten Reiners. 2015. Gami�cation. In Encyclopedia ofInformation Science and Technology. 3039–3047. https://doi.org/10.4018/978-1-4666-5888-2.ch297 01.

Session 3: VR + Games and Learning

181